Anthropic Introduces Managed Agents for Future-Proof AI Workflows
Anthropic launches Managed Agents, a new hosted service on the Claude Platform designed to run long-horizon AI agents through stable, virtualized interfaces. The system allows underlying implementations to change freely without breaking existing agent setups.
Anthropic introduces Managed Agents, a hosted service on the Claude Platform that runs long-horizon AI agents on behalf of developers. The engineering team highlights that traditional agent harnesses often encode assumptions about model limitations, such as Claude Sonnet 4.5's "context anxiety," which unexpectedly disappear in newer models like Claude Opus 4.5. This constant evolution makes rigid, tightly coupled agent setups quickly obsolete.
To solve this, Anthropic applies classic operating system design principles by virtualizing agent components into stable abstractions. The system breaks down the AI workflow into three distinct parts: a session for the append-only log, a harness for routing tool calls, and a sandbox for executing code. These interfaces remain consistent even as the underlying technology and model implementations change freely beneath them.
The development process reveals that Anthropic initially combined all components into a single container for simplicity but quickly encountered classic infrastructure problems. This tightly coupled "pet" approach creates unnecessary friction, reinforcing the team's decision to pivot toward modular, abstracted interfaces. Developers can now explore this new architecture by following the official Claude documentation.